Opaque Epistemic Mediation: The Unstable Truth of AI Models

TL;DR

• AI models provide inconsistent stances on fringe scientific claims, with answers varying significantly between web interfaces and developer APIs.
• This lack of stability turns AI into an opaque filter for knowledge, where the truth is determined by software updates rather than scientific consensus.

Background

As Large Language Models (LLMs) replace traditional search engines, they have become the primary gatekeepers of information for millions of users. Unlike a list of search results that allows a human to compare different viewpoints, an AI typically provides a single, authoritative-sounding answer. This shift places an immense responsibility on AI developers to ensure their models handle contested or pseudo-scientific claims with accuracy and consistency. However, the internal logic that governs these responses remains largely hidden from the public.

What happened

A new study has uncovered significant inconsistencies in how major AI models evaluate fringe scientific theories. Researchers tracked four prominent model families—Claude, Grok, GPT, and Gemini—over a five-month period from late 2025 to early 2026. They specifically tested the models' responses to ethnonationalist pseudo-science based on Frank Salter’s biosocial framework, a set of controversial theories often used to justify fringe political ideologies [^1]. The goal was to see if the models would consistently identify these theories as pseudo-science or if their stances would shift over time.

The findings revealed that LLM stances are neither stable nor transparent. The researchers identified a phenomenon they call opaque epistemic mediation, where the configuration of the model—whether it is accessed via a web chat interface or a raw API—changes the nature of the information provided. For instance, a model might correctly label a theory as pseudo-science when queried through its official website, which uses strict system prompts and safety filters. However, the same model might provide a neutral or even validating summary when accessed via an API, where those specific filters are absent or differently configured.

Furthermore, the study documented temporal drift, where a model's response to the same prompt changed between the four snapshots taken over five months. A model that was critical of a fringe theory in October might become more permissive or ambiguous by February. This aligns with previous research which found that the behavior of models like GPT-4 fluctuates significantly as companies tweak underlying parameters and training data [^2]. These changes often happen without any notification to the user, meaning the scientific consensus offered by the AI is in a state of constant, unannounced flux.

The study also noted differences between model families. While some models, like Claude and Gemini, tended to be more consistently aligned with mainstream scientific views due to rigorous safety tuning, others like Grok exhibited a more permissive stance toward contested claims. This creates a fragmented information environment where the choice of AI provider determines which fringe theories are validated and which are dismissed.

Why it matters

This research highlights a fundamental shift in how we verify facts. When the validation of a scientific claim depends on a software deployment configuration, the concept of a universal truth begins to erode. If a student uses a web-based AI to learn about biology while a researcher uses an API-based tool, they may receive conflicting information about what constitutes valid science. This discrepancy is not a result of new scientific discovery, but of opaque engineering decisions.

The danger of opaque epistemic mediation is that it allows for the quiet normalization of pseudo-science. If a model's safety filters are relaxed in an API version to improve performance or reduce refusals, fringe theories can slip through the cracks. Because users trust these models to be objective, a neutral summary of a debunked theory can be mistaken for a legitimate scientific debate. This is particularly concerning for biosocial frameworks that are used to support discriminatory ideologies. When an AI fails to provide a clear debunking, it effectively lends its perceived authority to the fringe claim.

Moreover, the lack of transparency prevents effective auditing. Because companies do not disclose the specific system prompts or reinforcement learning guidelines that shape these responses, it is impossible for the scientific community to know why a model's stance changed. We are moving toward a future where truth is a subscription service, and the quality of that truth depends on which tier or interface you pay for. This undermines the ideal of shared, verifiable facts. It turns scientific validation into a software setting that can be toggled by a handful of engineers behind closed doors.

Finally, this instability poses a challenge for developers building on top of AI. If an educational app relies on an LLM to provide factual information, and that LLM's stance on a topic drifts over three months, the app's reliability is compromised. We need stable standards for epistemic consistency in AI, ensuring that models provide evidence-based answers regardless of when or how they are queried. Without this, the AI becomes a mirror for the developer's current configuration rather than a window into reality.

Practical example

Imagine a college student named Sarah who is researching genetic interests for a sociology paper on a Tuesday morning. She asks a popular web-based AI for a summary. The web interface, equipped with a safety system prompt, tells her the theory is widely rejected by mainstream science. Sarah correctly notes this in her draft.

Meanwhile, her classmate, James, uses a specialized research tool that connects to the same AI via an API. Because the API version lacks the web-based safety layer, it provides James with a detailed, five-paragraph explanation of the theory without mentioning that it is considered pseudo-science. James concludes the theory is a valid area of study. When they meet to compare notes, they are confused. They used the same AI but received opposite conclusions. Neither realizes that a hidden software configuration, not the actual science, dictated their education.

Related gear

We recommend this book because it provides the critical thinking tools needed to navigate the exact type of data-driven pseudo-science and algorithmic bias explored in this study.